secureagentics/Adrian

What it solves

Adrian provides a runtime security layer for AI agents to prevent malicious or misaligned behavior. It addresses the limitation of static analysis and network monitoring by detecting attacks that slip past these traditional methods, such as prompt injections, jailbreaks, tool poisoning, data exfiltration, and privilege escalation.

How it works

Unlike traditional monitoring that only looks at activity logs (like tool calls and API interactions), Adrian analyzes both the agent's actual actions and its internal reasoning traces. By understanding why an agent is taking an action and correlating this across a session, it can identify when an agent goes "out-of-remit" (e.g., an e-commerce agent resetting passwords) even if the behavior hasn't been seen in a training set.

Who it’s for

Developers building AI agents using frameworks like LangChain, LangGraph, or Anthropic's SDK, as well as users of Claude Code. It is designed for those who need to ensure their agents operate within a defined remity and can be deployed via a managed cloud dashboard or entirely self-hosted for data sovereignty.

Highlights

• Reasoning-Aware Monitoring: Combines behavior and reasoning analysis to catch more nuanced attacks than behavior-only monitoring.
• In-flight Intervention: Ability to block or hold risky actions for human approval before they are executed.
• Broad Integration: Supports Python and TypeScript SDKs, with native plugins for Claude Code and support for LangChain.
• Self-Hosting Capability: Full stack deployment via Docker, including a local classifier model (Gemma 4) via Llama.cpp for offline, data-sovereign environments.